Most small companies already own the raw material for a useful AI assistant: a staff handbook, price lists, past proposals, service manuals and years of customer emails. What stops them is a reasonable fear that typing that material into a chatbot means handing it to someone else.
That fear is half right. The model itself is rarely where company data leaks. Leaks happen in the account someone opens with a personal email address, in a search index that ignores who is allowed to see what, and in chat logs nobody remembers to delete. Fix those three and a private assistant is well within reach of a small business.
What private AI for business actually means
Private AI for business is an assistant that answers questions from your company's own information while that information stays under contracts and access controls you choose. In most builds, the model is not retrained on your files at all. The system searches your documents when a question arrives, hands the most relevant passages to the model, and the model answers from them with a citation. The pattern is called retrieval-augmented generation, or RAG.
RAG matters for privacy for three practical reasons:
- You can remove a document. Take a file out of the index and it stops shaping answers. Knowledge baked into a model through training is far harder to withdraw.
- You can enforce permissions. The search step can be limited to what the person asking is allowed to open.
- You can check the answer. Every reply points back to a source a human can read.
Fine-tuning a model on your data has its uses, mostly for tone and format. It is rarely the right first step for a knowledge assistant, and it creates a copy of your information that is hard to audit.
Where company data actually leaks
Consumer accounts. The same AI brand can follow different rules depending on the plan. In an update announced on August 28, 2025, Anthropic let users of its Free, Pro and Max plans choose whether their chats help train its models, with data kept for five years for those who opt in. Its commercial terms, including the API, were excluded from that change. If staff paste client files into personal accounts, your own policy, not a vendor's business terms, is what protects you.
API and business defaults. OpenAI states that since March 1, 2023, data sent to its API has not been used to train its models unless the customer opts in, and that abuse monitoring logs are kept for up to 30 days by default, with zero data retention available only on approval. Anthropic says inputs and outputs from its commercial products are not used for training by default. Read the terms for the exact product you buy, because they differ.
Retrieval without permissions. OWASP's 2025 list of risks for language model applications warns that weak access controls on the vector stores behind RAG can expose sensitive information. In plain terms: if the payroll folder sits in the same index as the sales playbook, a sales rep can ask the assistant about salaries.
Prompt injection. OWASP ranks prompt injection first on the same list. A document or email can contain instructions aimed at the model. Keep any assistant that reads outside content away from tools that send, delete or pay.
Copies. Transcripts, search indexes, analytics events and backups are all copies of your data. Each one needs a home and a deletion date.
Three ways to run a private assistant
| Option | Where your data goes | Best for | Watch for |
|---|---|---|---|
| Business or API account with a major model provider | The provider's cloud, under business terms with no training by default | Most small teams | Retention windows, region options, staff on personal accounts |
| Models inside your own cloud account | Your cloud tenancy; AWS says model providers cannot see Amazon Bedrock prompts or completions | Companies already running on that cloud | Setup effort, usage costs, access policies |
| Self-hosted open-weight model | Your own servers | Strict confidentiality or offline needs | Hardware, updates, security and model quality are all on you |
For most owner-run companies, the first or second row is the right answer. Self-hosting buys control at the price of running infrastructure, and every leak point in the previous section still applies.
How to build one in six steps
- Pick one job. For example, answering staff questions from the handbook and price list, or drafting first replies from past proposals. One job is easier to test than a general assistant.
- Clean the sources. Remove outdated versions, merge duplicates and label sensitive folders such as HR, payroll and client contracts.
- Choose the account and read the terms. Confirm the training default, retention period, data region and sub-processors in writing, and sign a data processing agreement where your contracts require one.
- Build retrieval with permissions. Index content by access group, so the assistant only retrieves what the person asking could already open in your file system. Make it cite its sources.
- Limit actions. Start read-only. Anything that sends, updates or deletes should need a human click.
- Log, review and delete. Keep transcripts in a system you control, review a sample each month for wrong answers, and delete on a schedule.
Questions to ask any vendor before you sign
- Is our data used to train models by default, and where does the contract say so?
- How long are prompts, outputs and logs kept, and can we shorten that?
- Which sub-processors touch the data, and in which regions?
- Can retrieval enforce our existing folder and role permissions?
- Can we export the index, prompts and logs if we leave?
- Who owns the code, the configuration and the accounts?
At OWNIT, the answer to the last question is always the client: the setup, the data and the accounts sit in your name. Our private AI assistants run on whichever model provider fits your privacy, accuracy and cost needs, rather than one vendor we resell.
When a private AI build isn't worth it
- A small document set and a small team. If your knowledge fits in a few dozen files, a business plan from a major AI provider with training switched off may be enough. Write a short usage policy and skip the custom build.
- Messy or contradictory documents. An assistant repeats contradictions with confidence. Fix the source material first.
- Regulated data without the right contracts. Health, financial and children's data carry extra obligations. Do not connect them until the vendor agreements your regulators expect are signed.
- Nobody to own it. Answers drift as policies and prices change. If no one reviews the assistant monthly, it will quietly become wrong.
If you are still deciding what kind of AI you need at all, our guide to agents, chatbots and workflows sets out the differences. When you are ready to scope one job, OWNIT offers a free 30-minute strategy call to map where a private assistant would save your team the most time.
Frequently asked questions
Does ChatGPT or Claude train on my company's data?
It depends on the account. OpenAI says data sent to its API has not been used for training since March 1, 2023, unless the customer opts in. Anthropic says its commercial products, including the API, are not used for training by default, while users of its consumer plans choose for themselves. Check the terms for the exact plan your staff actually use.
What is RAG, and why does it help keep data private?
Retrieval-augmented generation, or RAG, means the assistant searches your documents when a question arrives and answers from the passages it finds, citing them. Your files are not baked into the model, so you can remove a document, restrict who can retrieve it and check every answer against its source. That makes access control and deletion far more practical.
Do we need to self-host an AI model to keep our data private?
Usually not. For most small businesses, a business or API account with no training by default, or a model running inside your own cloud account, gives strong protection with far less work. Self-hosting suits strict confidentiality or offline needs, but you then own the hardware, updates and security, and the main leak points, such as permissions and logs, remain.
How long does it take to build a private AI assistant?
A focused assistant built around one job, such as answering staff questions from your handbook and price lists, is often live within a few weeks once the source documents are cleaned and access is granted. Broader assistants that span several departments take longer, mainly because permissions and document cleanup take longer, not the AI itself.
Sources
Written by The OWNIT Editorial Desk and reviewed against public sources. General information, not legal or financial advice. Spotted an error? Email hello@ownit24.com and we will correct it.
